You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.

607 lines
26 KiB

//
// Copyright (c) 2017 The Khronos Group Inc.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
//
#include "harness/compat.h"
#include "harness/testHarness.h"
#include <string.h>
#include "cl_utils.h"
#include "tests.h"
#include <CL/cl_half.h>
int Test_vLoadHalf_private( cl_device_id device, bool aligned )
{
cl_int error;
int vectorSize;
cl_program programs[kVectorSizeCount+kStrangeVectorSizeCount][AS_NumAddressSpaces] = {{0}};
cl_kernel kernels[kVectorSizeCount+kStrangeVectorSizeCount][AS_NumAddressSpaces] = {{0}};
uint64_t time[kVectorSizeCount+kStrangeVectorSizeCount] = {0};
uint64_t min_time[kVectorSizeCount+kStrangeVectorSizeCount] = {0};
size_t q;
memset( min_time, -1, sizeof( min_time ) );
const char *vector_size_names[] = {"1", "2", "4", "8", "16", "3"};
int minVectorSize = kMinVectorSize;
// There is no aligned scalar vloada_half in CL 1.1
#if ! defined( CL_VERSION_1_1 ) && ! defined(__APPLE__)
vlog("Note: testing vloada_half.\n");
if (aligned && minVectorSize == 0)
minVectorSize = 1;
#endif
for( vectorSize = minVectorSize; vectorSize < kLastVectorSizeToTest; vectorSize++)
{
int effectiveVectorSize = g_arrVecSizes[vectorSize];
if(effectiveVectorSize == 3 && aligned) {
effectiveVectorSize = 4;
}
const char *source[] = {
"__kernel void test( const __global half *p, __global float", vector_size_name_extensions[vectorSize], " *f )\n"
"{\n"
" size_t i = get_global_id(0);\n"
" f[i] = vload", aligned ? "a" : "", "_half",vector_size_name_extensions[vectorSize],"( i, p );\n"
"}\n"
};
const char *sourceV3[] = {
"__kernel void test( const __global half *p, __global float *f,\n"
" uint extra_last_thread)\n"
"{\n"
" size_t i = get_global_id(0);\n"
" size_t last_i = get_global_size(0)-1;\n"
" if(last_i == i && extra_last_thread != 0) {\n"
" if(extra_last_thread ==2) {\n"
" f[3*i+1] = vload_half(3*i+1, p);\n"
" }\n"
" f[3*i] = vload_half(3*i, p);\n"
" } else {\n"
" vstore3(vload_half3( i, p ),i,f);\n"
" }\n"
"}\n"
};
const char *sourceV3aligned[] = {
"__kernel void test( const __global half *p, __global float3 *f )\n"
"{\n"
" size_t i = get_global_id(0);\n"
" f[i] = vloada_half3( i, p );\n"
" ((__global float *)f)[4*i+3] = vloada_half(4*i+3,p);\n"
"}\n"
};
const char *source_private1[] = {
"__kernel void test( const __global half *p, __global float *f )\n"
"{\n"
" __private ushort data[1];\n"
" __private half* hdata_p = (__private half*) data;\n"
" size_t i = get_global_id(0);\n"
" data[0] = ((__global ushort*)p)[i];\n"
" f[i] = vload", (aligned ? "a" : ""), "_half( 0, hdata_p );\n"
"}\n"
};
const char *source_private2[] = {
"__kernel void test( const __global half *p, __global float", vector_size_name_extensions[vectorSize], " *f )\n"
"{\n"
" __private ", align_types[vectorSize], " data[", vector_size_names[vectorSize], "/", align_divisors[vectorSize], "];\n"
" __private half* hdata_p = (__private half*) data;\n"
" __global ", align_types[vectorSize], "* i_p = (__global ", align_types[vectorSize], "*)p;\n"
" size_t i = get_global_id(0);\n"
" int k;\n"
" for (k=0; k<",vector_size_names[vectorSize],"/",align_divisors[vectorSize],"; k++)\n"
" data[k] = i_p[i+k];\n"
" f[i] = vload", aligned ? "a" : "", "_half",vector_size_name_extensions[vectorSize],"( 0, hdata_p );\n"
"}\n"
};
const char *source_privateV3[] = {
"__kernel void test( const __global half *p, __global float *f,"
" uint extra_last_thread )\n"
"{\n"
" __private ushort data[3];\n"
" __private half* hdata_p = (__private half*) data;\n"
" __global ushort* i_p = (__global ushort*)p;\n"
" size_t i = get_global_id(0);\n"
" int k;\n"
// " data = vload3(i, i_p);\n"
" size_t last_i = get_global_size(0)-1;\n"
" if(last_i == i && extra_last_thread != 0) {\n"
" if(extra_last_thread ==2) {\n"
" f[3*i+1] = vload_half(3*i+1, p);\n"
" }\n"
" f[3*i] = vload_half(3*i, p);\n"
" } else {\n"
" for (k=0; k<3; k++)\n"
" data[k] = i_p[i*3+k];\n"
" vstore3(vload_half3( 0, hdata_p ), i, f);\n"
" }\n"
"}\n"
};
const char *source_privateV3aligned[] = {
"__kernel void test( const __global half *p, __global float3 *f )\n"
"{\n"
" ushort4 data[4];\n" // declare as vector for alignment. Make four to check to see vloada_half3 index is working.
" half* hdata_p = (half*) &data;\n"
" size_t i = get_global_id(0);\n"
" global ushort* i_p = (global ushort*)p + i * 4;\n"
" int offset = i & 3;\n"
" data[offset] = (ushort4)( i_p[0], i_p[1], i_p[2], USHRT_MAX ); \n"
" data[offset^1] = USHRT_MAX; \n"
" data[offset^2] = USHRT_MAX; \n"
" data[offset^3] = USHRT_MAX; \n"
// test vloada_half3
" f[i] = vloada_half3( offset, hdata_p );\n"
// Fill in the 4th value so we don't have to special case this code elsewhere in the test.
" mem_fence(CLK_GLOBAL_MEM_FENCE );\n"
" ((__global float *)f)[4*i+3] = vload_half(4*i+3, p);\n"
"}\n"
};
char local_buf_size[10];
sprintf(local_buf_size, "%lld", (uint64_t)((effectiveVectorSize))*gWorkGroupSize);
const char *source_local1[] = {
"__kernel void test( const __global half *p, __global float *f )\n"
"{\n"
" __local ushort data[",local_buf_size,"];\n"
" __local half* hdata_p = (__local half*) data;\n"
" size_t i = get_global_id(0);\n"
" size_t lid = get_local_id(0);\n"
" data[lid] = ((__global ushort*)p)[i];\n"
" f[i] = vload", aligned ? "a" : "", "_half( lid, hdata_p );\n"
"}\n"
};
const char *source_local2[] = {
"#define VECTOR_LEN (",
vector_size_names[vectorSize],
"/",
align_divisors[vectorSize],
")\n"
"#define ALIGN_TYPE ",
align_types[vectorSize],
"\n"
"__kernel void test( const __global half *p, __global float",
vector_size_name_extensions[vectorSize],
" *f )\n"
"{\n"
" __local uchar data[",
local_buf_size,
"/",
align_divisors[vectorSize],
"*sizeof(ALIGN_TYPE)] ",
"__attribute__((aligned(sizeof(ALIGN_TYPE))));\n"
" __local half* hdata_p = (__local half*) data;\n"
" __global ALIGN_TYPE* i_p = (__global ALIGN_TYPE*)p;\n"
" size_t i = get_global_id(0);\n"
" size_t lid = get_local_id(0);\n"
" int k;\n"
" for (k=0; k<VECTOR_LEN; k++)\n"
" *(__local ",
"ALIGN_TYPE*)&(data[(lid*VECTOR_LEN+k)*sizeof(ALIGN_TYPE)]) = ",
"i_p[i*VECTOR_LEN+k];\n"
" f[i] = vload",
aligned ? "a" : "",
"_half",
vector_size_name_extensions[vectorSize],
"( lid, hdata_p );\n"
"}\n"
};
const char *source_localV3[] = {
"__kernel void test( const __global half *p, __global float *f,\n"
" uint extra_last_thread)\n"
"{\n"
" __local ushort data[", local_buf_size,"];\n"
" __local half* hdata_p = (__local half*) data;\n"
" __global ushort* i_p = (__global ushort*)p;\n"
" size_t i = get_global_id(0);\n"
" size_t last_i = get_global_size(0)-1;\n"
" size_t lid = get_local_id(0);\n"
" int k;\n"
" if(last_i == i && extra_last_thread != 0) {\n"
" if(extra_last_thread ==2) {\n"
" f[3*i+1] = vload_half(3*i+1, p);\n"
" }\n"
" f[3*i] = vload_half(3*i, p);\n"
" } else {\n"
" for (k=0; k<3; k++)\n"
" data[lid*3+k] = i_p[i*3+k];\n"
" vstore3( vload_half3( lid, hdata_p ),i,f);\n"
" };\n"
"}\n"
};
const char *source_localV3aligned[] = {
"__kernel void test( const __global half *p, __global float3 *f )\n"
"{\n"
" __local ushort data[", local_buf_size,"];\n"
" __local half* hdata_p = (__local half*) data;\n"
" __global ushort* i_p = (__global ushort*)p;\n"
" size_t i = get_global_id(0);\n"
" size_t lid = get_local_id(0);\n"
" int k;\n"
" for (k=0; k<4; k++)\n"
" data[lid*4+k] = i_p[i*4+k];\n"
" f[i] = vloada_half3( lid, hdata_p );\n"
" ((__global float *)f)[4*i+3] = vload_half(lid*4+3, hdata_p);\n"
"}\n"
};
const char *source_constant[] = {
"__kernel void test( __constant half *p, __global float", vector_size_name_extensions[vectorSize], " *f )\n"
"{\n"
" size_t i = get_global_id(0);\n"
" f[i] = vload", aligned ? "a" : "", "_half",vector_size_name_extensions[vectorSize],"( i, p );\n"
"}\n"
};
const char *source_constantV3[] = {
"__kernel void test( __constant half *p, __global float *f,\n"
" uint extra_last_thread)\n"
"{\n"
" size_t i = get_global_id(0);\n"
" size_t last_i = get_global_size(0)-1;\n"
" if(last_i == i && extra_last_thread != 0) {\n"
" if(extra_last_thread ==2) {\n"
" f[3*i+1] = vload_half(3*i+1, p);\n"
" }\n"
" f[3*i] = vload_half(3*i, p);\n"
" } else {\n"
" vstore3(vload_half",vector_size_name_extensions[vectorSize],"( i, p ), i, f);\n"
" }\n"
"}\n"
};
const char *source_constantV3aligned[] = {
"__kernel void test( __constant half *p, __global float3 *f )\n"
"{\n"
" size_t i = get_global_id(0);\n"
" f[i] = vloada_half3( i, p );\n"
" ((__global float *)f)[4*i+3] = vload_half(4*i+3,p);\n"
"}\n"
};
if(g_arrVecSizes[vectorSize] != 3) {
programs[vectorSize][AS_Global] = MakeProgram( device, source, sizeof( source) / sizeof( source[0]) );
if( NULL == programs[ vectorSize ][AS_Global] ) {
gFailCount++;
vlog_error( "\t\tFAILED -- Failed to create program.\n" );
for ( q= 0; q < sizeof( source) / sizeof( source[0]); q++)
vlog_error("%s", source[q]);
return -1;
} else {
}
} else if(aligned) {
programs[vectorSize][AS_Global] = MakeProgram( device, sourceV3aligned, sizeof( sourceV3aligned) / sizeof( sourceV3aligned[0]) );
if( NULL == programs[ vectorSize ][AS_Global] ) {
gFailCount++;
vlog_error( "\t\tFAILED -- Failed to create program.\n" );
for ( q= 0; q < sizeof( sourceV3aligned) / sizeof( sourceV3aligned[0]); q++)
vlog_error("%s", sourceV3aligned[q]);
return -1;
} else {
}
} else {
programs[vectorSize][AS_Global] = MakeProgram( device, sourceV3, sizeof( sourceV3) / sizeof( sourceV3[0]) );
if( NULL == programs[ vectorSize ][AS_Global] ) {
gFailCount++;
vlog_error( "\t\tFAILED -- Failed to create program.\n" );
for ( q= 0; q < sizeof( sourceV3) / sizeof( sourceV3[0]); q++)
vlog_error("%s", sourceV3[q]);
return -1;
}
}
kernels[ vectorSize ][AS_Global] = clCreateKernel( programs[ vectorSize ][AS_Global], "test", &error );
if( NULL == kernels[vectorSize][AS_Global] )
{
gFailCount++;
vlog_error( "\t\tFAILED -- Failed to create kernel. (%d)\n", error );
return -2;
}
const char** source_ptr;
uint32_t source_size;
if (vectorSize == 0) {
source_ptr = source_private1;
source_size = sizeof( source_private1) / sizeof( source_private1[0]);
} else if(g_arrVecSizes[vectorSize] == 3) {
if(aligned) {
source_ptr = source_privateV3aligned;
source_size = sizeof( source_privateV3aligned) / sizeof( source_privateV3aligned[0]);
} else {
source_ptr = source_privateV3;
source_size = sizeof( source_privateV3) / sizeof( source_privateV3[0]);
}
} else {
source_ptr = source_private2;
source_size = sizeof( source_private2) / sizeof( source_private2[0]);
}
programs[vectorSize][AS_Private] = MakeProgram( device, source_ptr, source_size );
if( NULL == programs[ vectorSize ][AS_Private] )
{
gFailCount++;
vlog_error( "\t\tFAILED -- Failed to create private program.\n" );
for ( q= 0; q < source_size; q++)
vlog_error("%s", source_ptr[q]);
return -1;
}
kernels[ vectorSize ][AS_Private] = clCreateKernel( programs[ vectorSize ][AS_Private], "test", &error );
if( NULL == kernels[vectorSize][AS_Private] )
{
gFailCount++;
vlog_error( "\t\tFAILED -- Failed to create private kernel. (%d)\n", error );
return -2;
}
if (vectorSize == 0) {
source_ptr = source_local1;
source_size = sizeof( source_local1) / sizeof( source_local1[0]);
} else if(g_arrVecSizes[vectorSize] == 3) {
if(aligned) {
source_ptr = source_localV3aligned;
source_size = sizeof(source_localV3aligned)/sizeof(source_localV3aligned[0]);
} else {
source_ptr = source_localV3;
source_size = sizeof(source_localV3)/sizeof(source_localV3[0]);
}
} else {
source_ptr = source_local2;
source_size = sizeof( source_local2) / sizeof( source_local2[0]);
}
programs[vectorSize][AS_Local] = MakeProgram( device, source_ptr, source_size );
if( NULL == programs[ vectorSize ][AS_Local] )
{
gFailCount++;
vlog_error( "\t\tFAILED -- Failed to create local program.\n" );
for ( q= 0; q < source_size; q++)
vlog_error("%s", source_ptr[q]);
return -1;
}
kernels[ vectorSize ][AS_Local] = clCreateKernel( programs[ vectorSize ][AS_Local], "test", &error );
if( NULL == kernels[vectorSize][AS_Local] )
{
gFailCount++;
vlog_error( "\t\tFAILED -- Failed to create local kernel. (%d)\n", error );
return -2;
}
if(g_arrVecSizes[vectorSize] == 3) {
if(aligned) {
programs[vectorSize][AS_Constant] = MakeProgram( device, source_constantV3aligned, sizeof(source_constantV3aligned) / sizeof( source_constantV3aligned[0]) );
if( NULL == programs[ vectorSize ][AS_Constant] )
{
gFailCount++;
vlog_error( "\t\tFAILED -- Failed to create constant program.\n" );
for ( q= 0; q < sizeof( source_constantV3aligned) / sizeof( source_constantV3aligned[0]); q++)
vlog_error("%s", source_constantV3aligned[q]);
return -1;
}
} else {
programs[vectorSize][AS_Constant] = MakeProgram( device, source_constantV3, sizeof(source_constantV3) / sizeof( source_constantV3[0]) );
if( NULL == programs[ vectorSize ][AS_Constant] )
{
gFailCount++;
vlog_error( "\t\tFAILED -- Failed to create constant program.\n" );
for ( q= 0; q < sizeof( source_constantV3) / sizeof( source_constantV3[0]); q++)
vlog_error("%s", source_constantV3[q]);
return -1;
}
}
} else {
programs[vectorSize][AS_Constant] = MakeProgram( device, source_constant, sizeof(source_constant) / sizeof( source_constant[0]) );
if( NULL == programs[ vectorSize ][AS_Constant] )
{
gFailCount++;
vlog_error( "\t\tFAILED -- Failed to create constant program.\n" );
for ( q= 0; q < sizeof( source_constant) / sizeof( source_constant[0]); q++)
vlog_error("%s", source_constant[q]);
return -1;
}
}
kernels[ vectorSize ][AS_Constant] = clCreateKernel( programs[ vectorSize ][AS_Constant], "test", &error );
if( NULL == kernels[vectorSize][AS_Constant] )
{
gFailCount++;
vlog_error( "\t\tFAILED -- Failed to create constant kernel. (%d)\n", error );
return -2;
}
}
// Figure out how many elements are in a work block
size_t elementSize = MAX( sizeof(cl_half), sizeof(cl_float));
size_t blockCount = getBufferSize(device) / elementSize; // elementSize is power of 2
uint64_t lastCase = 1ULL << (8*sizeof(cl_half)); // number of things of size cl_half
// we handle 64-bit types a bit differently.
if( lastCase == 0 )
lastCase = 0x100000000ULL;
uint64_t i, j;
uint64_t printMask = (lastCase >> 4) - 1;
uint32_t count = 0;
error = 0;
int addressSpace;
// int reported_vector_skip = 0;
for( i = 0; i < (uint64_t)lastCase; i += blockCount )
{
count = (uint32_t) MIN( blockCount, lastCase - i );
//Init the input stream
uint16_t *p = (uint16_t *)gIn_half;
for( j = 0; j < count; j++ )
p[j] = j + i;
if( (error = clEnqueueWriteBuffer(gQueue, gInBuffer_half, CL_TRUE, 0, count * sizeof( cl_half ), gIn_half, 0, NULL, NULL)))
{
vlog_error( "Failure in clWriteArray\n" );
gFailCount++;
goto exit;
}
//create the reference result
const unsigned short *s = (const unsigned short *)gIn_half;
float *d = (float *)gOut_single_reference;
for (j = 0; j < count; j++) d[j] = cl_half_to_float(s[j]);
//Check the vector lengths
for( vectorSize = minVectorSize; vectorSize < kLastVectorSizeToTest; vectorSize++)
{ // here we loop through vector sizes, 3 is last
for ( addressSpace = 0; addressSpace < AS_NumAddressSpaces; addressSpace++) {
uint32_t pattern = 0x7fffdead;
/*
if (addressSpace == 3) {
vlog("Note: skipping address space %s due to small buffer size.\n", addressSpaceNames[addressSpace]);
continue;
}
*/
memset_pattern4( gOut_single, &pattern, getBufferSize(device));
if( (error = clEnqueueWriteBuffer(gQueue, gOutBuffer_single, CL_TRUE, 0, count * sizeof( float ), gOut_single, 0, NULL, NULL)) )
{
vlog_error( "Failure in clWriteArray\n" );
gFailCount++;
goto exit;
}
if(g_arrVecSizes[vectorSize] == 3 && !aligned) {
// now we need to add the extra const argument for how
// many elements the last thread should take care of.
}
// okay, here is where we have to be careful
if( (error = RunKernel(device, kernels[vectorSize][addressSpace], gInBuffer_half, gOutBuffer_single, numVecs(count, vectorSize, aligned) ,
runsOverBy(count, vectorSize, aligned) ) ) )
{
gFailCount++;
goto exit;
}
if( (error = clEnqueueReadBuffer(gQueue, gOutBuffer_single, CL_TRUE, 0, count * sizeof( float ), gOut_single, 0, NULL, NULL)) )
{
vlog_error( "Failure in clReadArray\n" );
gFailCount++;
goto exit;
}
if( memcmp( gOut_single, gOut_single_reference, count * sizeof( float )) )
{
uint32_t *u1 = (uint32_t *)gOut_single;
uint32_t *u2 = (uint32_t *)gOut_single_reference;
float *f1 = (float *)gOut_single;
float *f2 = (float *)gOut_single_reference;
for( j = 0; j < count; j++ )
{
if(isnan(f1[j]) && isnan(f2[j])) // both are nan dont compare them
continue;
if( u1[j] != u2[j])
{
vlog_error( " %lld) (of %lld) Failure at 0x%4.4x: %a vs *%a (0x%8.8x vs *0x%8.8x) vector_size = %d (%s) address space = %s, load is %s\n",
j, (uint64_t)count, ((unsigned short*)gIn_half)[j], f1[j], f2[j], u1[j], u2[j], (g_arrVecSizes[vectorSize]),
vector_size_names[vectorSize], addressSpaceNames[addressSpace],
(aligned?"aligned":"unaligned"));
gFailCount++;
error = -1;
goto exit;
}
}
}
if( gReportTimes && addressSpace == 0)
{
//Run again for timing
for( j = 0; j < 100; j++ )
{
uint64_t startTime = ReadTime();
error =
RunKernel(device, kernels[vectorSize][addressSpace], gInBuffer_half, gOutBuffer_single, numVecs(count, vectorSize, aligned) ,
runsOverBy(count, vectorSize, aligned));
if(error)
{
gFailCount++;
goto exit;
}
if( (error = clFinish(gQueue)) )
{
vlog_error( "Failure in clFinish\n" );
gFailCount++;
goto exit;
}
uint64_t currentTime = ReadTime() - startTime;
time[ vectorSize ] += currentTime;
if( currentTime < min_time[ vectorSize ] )
min_time[ vectorSize ] = currentTime ;
}
}
}
}
if( ((i+blockCount) & ~printMask) == (i+blockCount) )
{
vlog( "." );
fflush( stdout );
}
}
vlog( "\n" );
if( gReportTimes )
{
for( vectorSize = minVectorSize; vectorSize < kLastVectorSizeToTest; vectorSize++)
vlog_perf( SubtractTime( time[ vectorSize ], 0 ) * 1e6 * gDeviceFrequency * gComputeDevices / (double) (count * 100), 0,
"average us/elem", "vLoad%sHalf avg. (%s, vector size: %d)", ( (aligned) ? "a" : ""), addressSpaceNames[0], (g_arrVecSizes[vectorSize]) );
for( vectorSize = minVectorSize; vectorSize < kLastVectorSizeToTest; vectorSize++)
vlog_perf( SubtractTime( min_time[ vectorSize ], 0 ) * 1e6 * gDeviceFrequency * gComputeDevices / (double) count, 0,
"best us/elem", "vLoad%sHalf best (%s vector size: %d)", ( (aligned) ? "a" : ""), addressSpaceNames[0], (g_arrVecSizes[vectorSize]) );
}
exit:
//clean up
for( vectorSize = minVectorSize; vectorSize < kLastVectorSizeToTest; vectorSize++)
{
for ( addressSpace = 0; addressSpace < AS_NumAddressSpaces; addressSpace++) {
clReleaseKernel( kernels[ vectorSize ][addressSpace] );
clReleaseProgram( programs[ vectorSize ][addressSpace] );
}
}
return error;
}
int test_vload_half( cl_device_id device, cl_context context, cl_command_queue queue, int num_elements )
{
return Test_vLoadHalf_private( device, false );
}
int test_vloada_half( cl_device_id device, cl_context context, cl_command_queue queue, int num_elements )
{
return Test_vLoadHalf_private( device, true );
}